2 * PowerPC64 port by Mike Corrigan and Dave Engebretsen
3 * {mikejc|engebret}@us.ibm.com
5 * Copyright (c) 2000 Mike Corrigan <mikejc@us.ibm.com>
7 * SMP scalability work:
8 * Copyright (C) 2001 Anton Blanchard <anton@au.ibm.com>, IBM
13 * PowerPC Hashed Page Table functions
15 * This program is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU General Public License
17 * as published by the Free Software Foundation; either version
18 * 2 of the License, or (at your option) any later version.
24 #include <linux/spinlock.h>
25 #include <linux/errno.h>
26 #include <linux/sched.h>
27 #include <linux/proc_fs.h>
28 #include <linux/stat.h>
29 #include <linux/sysctl.h>
30 #include <linux/export.h>
31 #include <linux/ctype.h>
32 #include <linux/cache.h>
33 #include <linux/init.h>
34 #include <linux/signal.h>
35 #include <linux/memblock.h>
36 #include <linux/context_tracking.h>
37 #include <linux/libfdt.h>
39 #include <asm/processor.h>
40 #include <asm/pgtable.h>
42 #include <asm/mmu_context.h>
44 #include <asm/types.h>
45 #include <asm/uaccess.h>
46 #include <asm/machdep.h>
48 #include <asm/tlbflush.h>
52 #include <asm/cacheflush.h>
53 #include <asm/cputable.h>
54 #include <asm/sections.h>
55 #include <asm/copro.h>
57 #include <asm/code-patching.h>
58 #include <asm/fadump.h>
59 #include <asm/firmware.h>
61 #include <asm/trace.h>
65 #define DBG(fmt...) udbg_printf(fmt)
71 #define DBG_LOW(fmt...) udbg_printf(fmt)
73 #define DBG_LOW(fmt...)
81 * Note: pte --> Linux PTE
82 * HPTE --> PowerPC Hashed Page Table Entry
85 * htab_initialize is called with the MMU off (of course), but
86 * the kernel has been copied down to zero so it can directly
87 * reference global data. At this point it is very difficult
88 * to print debug info.
92 static unsigned long _SDR1;
93 struct mmu_psize_def mmu_psize_defs[MMU_PAGE_COUNT];
94 EXPORT_SYMBOL_GPL(mmu_psize_defs);
96 u8 hpte_page_sizes[1 << LP_BITS];
97 EXPORT_SYMBOL_GPL(hpte_page_sizes);
99 struct hash_pte *htab_address;
100 unsigned long htab_size_bytes;
101 unsigned long htab_hash_mask;
102 EXPORT_SYMBOL_GPL(htab_hash_mask);
103 int mmu_linear_psize = MMU_PAGE_4K;
104 EXPORT_SYMBOL_GPL(mmu_linear_psize);
105 int mmu_virtual_psize = MMU_PAGE_4K;
106 int mmu_vmalloc_psize = MMU_PAGE_4K;
107 #ifdef CONFIG_SPARSEMEM_VMEMMAP
108 int mmu_vmemmap_psize = MMU_PAGE_4K;
110 int mmu_io_psize = MMU_PAGE_4K;
111 int mmu_kernel_ssize = MMU_SEGSIZE_256M;
112 EXPORT_SYMBOL_GPL(mmu_kernel_ssize);
113 int mmu_highuser_ssize = MMU_SEGSIZE_256M;
114 u16 mmu_slb_size = 64;
115 EXPORT_SYMBOL_GPL(mmu_slb_size);
116 #ifdef CONFIG_PPC_64K_PAGES
117 int mmu_ci_restrictions;
119 #ifdef CONFIG_DEBUG_PAGEALLOC
120 static u8 *linear_map_hash_slots;
121 static unsigned long linear_map_hash_count;
122 static DEFINE_SPINLOCK(linear_map_hash_lock);
123 #endif /* CONFIG_DEBUG_PAGEALLOC */
124 struct mmu_hash_ops mmu_hash_ops;
125 EXPORT_SYMBOL(mmu_hash_ops);
127 /* There are definitions of page sizes arrays to be used when none
128 * is provided by the firmware.
131 /* Pre-POWER4 CPUs (4k pages only)
133 static struct mmu_psize_def mmu_psize_defaults_old[] = {
137 .penc = {[MMU_PAGE_4K] = 0, [1 ... MMU_PAGE_COUNT - 1] = -1},
143 /* POWER4, GPUL, POWER5
145 * Support for 16Mb large pages
147 static struct mmu_psize_def mmu_psize_defaults_gp[] = {
151 .penc = {[MMU_PAGE_4K] = 0, [1 ... MMU_PAGE_COUNT - 1] = -1},
158 .penc = {[0 ... MMU_PAGE_16M - 1] = -1, [MMU_PAGE_16M] = 0,
159 [MMU_PAGE_16M + 1 ... MMU_PAGE_COUNT - 1] = -1 },
166 * 'R' and 'C' update notes:
167 * - Under pHyp or KVM, the updatepp path will not set C, thus it *will*
168 * create writeable HPTEs without C set, because the hcall H_PROTECT
169 * that we use in that case will not update C
170 * - The above is however not a problem, because we also don't do that
171 * fancy "no flush" variant of eviction and we use H_REMOVE which will
172 * do the right thing and thus we don't have the race I described earlier
174 * - Under bare metal, we do have the race, so we need R and C set
175 * - We make sure R is always set and never lost
176 * - C is _PAGE_DIRTY, and *should* always be set for a writeable mapping
178 unsigned long htab_convert_pte_flags(unsigned long pteflags)
180 unsigned long rflags = 0;
182 /* _PAGE_EXEC -> NOEXEC */
183 if ((pteflags & _PAGE_EXEC) == 0)
187 * Linux uses slb key 0 for kernel and 1 for user.
188 * kernel RW areas are mapped with PPP=0b000
189 * User area is mapped with PPP=0b010 for read/write
190 * or PPP=0b011 for read-only (including writeable but clean pages).
192 if (pteflags & _PAGE_PRIVILEGED) {
194 * Kernel read only mapped with ppp bits 0b110
196 if (!(pteflags & _PAGE_WRITE)) {
197 if (mmu_has_feature(MMU_FTR_KERNEL_RO))
198 rflags |= (HPTE_R_PP0 | 0x2);
203 if (pteflags & _PAGE_RWX)
205 if (!((pteflags & _PAGE_WRITE) && (pteflags & _PAGE_DIRTY)))
209 * We can't allow hardware to update hpte bits. Hence always
210 * set 'R' bit and set 'C' if it is a write fault
214 if (pteflags & _PAGE_DIRTY)
220 if ((pteflags & _PAGE_CACHE_CTL) == _PAGE_TOLERANT)
222 else if ((pteflags & _PAGE_CACHE_CTL) == _PAGE_NON_IDEMPOTENT)
223 rflags |= (HPTE_R_I | HPTE_R_G);
224 else if ((pteflags & _PAGE_CACHE_CTL) == _PAGE_SAO)
225 rflags |= (HPTE_R_W | HPTE_R_I | HPTE_R_M);
228 * Add memory coherence if cache inhibited is not set
235 int htab_bolt_mapping(unsigned long vstart, unsigned long vend,
236 unsigned long pstart, unsigned long prot,
237 int psize, int ssize)
239 unsigned long vaddr, paddr;
240 unsigned int step, shift;
243 shift = mmu_psize_defs[psize].shift;
246 prot = htab_convert_pte_flags(prot);
248 DBG("htab_bolt_mapping(%lx..%lx -> %lx (%lx,%d,%d)\n",
249 vstart, vend, pstart, prot, psize, ssize);
251 for (vaddr = vstart, paddr = pstart; vaddr < vend;
252 vaddr += step, paddr += step) {
253 unsigned long hash, hpteg;
254 unsigned long vsid = get_kernel_vsid(vaddr, ssize);
255 unsigned long vpn = hpt_vpn(vaddr, vsid, ssize);
256 unsigned long tprot = prot;
259 * If we hit a bad address return error.
263 /* Make kernel text executable */
264 if (overlaps_kernel_text(vaddr, vaddr + step))
267 /* Make kvm guest trampolines executable */
268 if (overlaps_kvm_tmp(vaddr, vaddr + step))
272 * If relocatable, check if it overlaps interrupt vectors that
273 * are copied down to real 0. For relocatable kernel
274 * (e.g. kdump case) we copy interrupt vectors down to real
275 * address 0. Mark that region as executable. This is
276 * because on p8 system with relocation on exception feature
277 * enabled, exceptions are raised with MMU (IR=DR=1) ON. Hence
278 * in order to execute the interrupt handlers in virtual
279 * mode the vector region need to be marked as executable.
281 if ((PHYSICAL_START > MEMORY_START) &&
282 overlaps_interrupt_vector_text(vaddr, vaddr + step))
285 hash = hpt_hash(vpn, shift, ssize);
286 hpteg = ((hash & htab_hash_mask) * HPTES_PER_GROUP);
288 BUG_ON(!mmu_hash_ops.hpte_insert);
289 ret = mmu_hash_ops.hpte_insert(hpteg, vpn, paddr, tprot,
290 HPTE_V_BOLTED, psize, psize,
293 /* Try to remove a non bolted entry */
294 ret = mmu_hash_ops.hpte_remove(hpteg);
296 ret = mmu_hash_ops.hpte_insert(hpteg, vpn, paddr, tprot,
297 HPTE_V_BOLTED, psize, psize,
304 #ifdef CONFIG_DEBUG_PAGEALLOC
305 if (debug_pagealloc_enabled() &&
306 (paddr >> PAGE_SHIFT) < linear_map_hash_count)
307 linear_map_hash_slots[paddr >> PAGE_SHIFT] = ret | 0x80;
308 #endif /* CONFIG_DEBUG_PAGEALLOC */
310 return ret < 0 ? ret : 0;
313 int htab_remove_mapping(unsigned long vstart, unsigned long vend,
314 int psize, int ssize)
317 unsigned int step, shift;
321 shift = mmu_psize_defs[psize].shift;
324 if (!mmu_hash_ops.hpte_removebolted)
327 for (vaddr = vstart; vaddr < vend; vaddr += step) {
328 rc = mmu_hash_ops.hpte_removebolted(vaddr, psize, ssize);
340 static bool disable_1tb_segments = false;
342 static int __init parse_disable_1tb_segments(char *p)
344 disable_1tb_segments = true;
347 early_param("disable_1tb_segments", parse_disable_1tb_segments);
349 static int __init htab_dt_scan_seg_sizes(unsigned long node,
350 const char *uname, int depth,
353 const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
357 /* We are scanning "cpu" nodes only */
358 if (type == NULL || strcmp(type, "cpu") != 0)
361 prop = of_get_flat_dt_prop(node, "ibm,processor-segment-sizes", &size);
364 for (; size >= 4; size -= 4, ++prop) {
365 if (be32_to_cpu(prop[0]) == 40) {
366 DBG("1T segment support detected\n");
368 if (disable_1tb_segments) {
369 DBG("1T segments disabled by command line\n");
373 cur_cpu_spec->mmu_features |= MMU_FTR_1T_SEGMENT;
377 cur_cpu_spec->mmu_features &= ~MMU_FTR_NO_SLBIE_B;
381 static int __init get_idx_from_shift(unsigned int shift)
405 static int __init htab_dt_scan_page_sizes(unsigned long node,
406 const char *uname, int depth,
409 const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
413 /* We are scanning "cpu" nodes only */
414 if (type == NULL || strcmp(type, "cpu") != 0)
417 prop = of_get_flat_dt_prop(node, "ibm,segment-page-sizes", &size);
421 pr_info("Page sizes from device-tree:\n");
423 cur_cpu_spec->mmu_features &= ~(MMU_FTR_16M_PAGE);
425 unsigned int base_shift = be32_to_cpu(prop[0]);
426 unsigned int slbenc = be32_to_cpu(prop[1]);
427 unsigned int lpnum = be32_to_cpu(prop[2]);
428 struct mmu_psize_def *def;
431 size -= 3; prop += 3;
432 base_idx = get_idx_from_shift(base_shift);
434 /* skip the pte encoding also */
435 prop += lpnum * 2; size -= lpnum * 2;
438 def = &mmu_psize_defs[base_idx];
439 if (base_idx == MMU_PAGE_16M)
440 cur_cpu_spec->mmu_features |= MMU_FTR_16M_PAGE;
442 def->shift = base_shift;
443 if (base_shift <= 23)
446 def->avpnm = (1 << (base_shift - 23)) - 1;
449 * We don't know for sure what's up with tlbiel, so
450 * for now we only set it for 4K and 64K pages
452 if (base_idx == MMU_PAGE_4K || base_idx == MMU_PAGE_64K)
457 while (size > 0 && lpnum) {
458 unsigned int shift = be32_to_cpu(prop[0]);
459 int penc = be32_to_cpu(prop[1]);
461 prop += 2; size -= 2;
464 idx = get_idx_from_shift(shift);
469 pr_err("Invalid penc for base_shift=%d "
470 "shift=%d\n", base_shift, shift);
472 def->penc[idx] = penc;
473 pr_info("base_shift=%d: shift=%d, sllp=0x%04lx,"
474 " avpnm=0x%08lx, tlbiel=%d, penc=%d\n",
475 base_shift, shift, def->sllp,
476 def->avpnm, def->tlbiel, def->penc[idx]);
483 #ifdef CONFIG_HUGETLB_PAGE
484 /* Scan for 16G memory blocks that have been set aside for huge pages
485 * and reserve those blocks for 16G huge pages.
487 static int __init htab_dt_scan_hugepage_blocks(unsigned long node,
488 const char *uname, int depth,
490 const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
491 const __be64 *addr_prop;
492 const __be32 *page_count_prop;
493 unsigned int expected_pages;
494 long unsigned int phys_addr;
495 long unsigned int block_size;
497 /* We are scanning "memory" nodes only */
498 if (type == NULL || strcmp(type, "memory") != 0)
501 /* This property is the log base 2 of the number of virtual pages that
502 * will represent this memory block. */
503 page_count_prop = of_get_flat_dt_prop(node, "ibm,expected#pages", NULL);
504 if (page_count_prop == NULL)
506 expected_pages = (1 << be32_to_cpu(page_count_prop[0]));
507 addr_prop = of_get_flat_dt_prop(node, "reg", NULL);
508 if (addr_prop == NULL)
510 phys_addr = be64_to_cpu(addr_prop[0]);
511 block_size = be64_to_cpu(addr_prop[1]);
512 if (block_size != (16 * GB))
514 printk(KERN_INFO "Huge page(16GB) memory: "
515 "addr = 0x%lX size = 0x%lX pages = %d\n",
516 phys_addr, block_size, expected_pages);
517 if (phys_addr + (16 * GB) <= memblock_end_of_DRAM()) {
518 memblock_reserve(phys_addr, block_size * expected_pages);
519 add_gpage(phys_addr, block_size, expected_pages);
523 #endif /* CONFIG_HUGETLB_PAGE */
525 static void mmu_psize_set_default_penc(void)
528 for (bpsize = 0; bpsize < MMU_PAGE_COUNT; bpsize++)
529 for (apsize = 0; apsize < MMU_PAGE_COUNT; apsize++)
530 mmu_psize_defs[bpsize].penc[apsize] = -1;
533 #ifdef CONFIG_PPC_64K_PAGES
535 static bool might_have_hea(void)
538 * The HEA ethernet adapter requires awareness of the
539 * GX bus. Without that awareness we can easily assume
540 * we will never see an HEA ethernet device.
542 #ifdef CONFIG_IBMEBUS
543 return !cpu_has_feature(CPU_FTR_ARCH_207S) &&
544 firmware_has_feature(FW_FEATURE_SPLPAR);
550 #endif /* #ifdef CONFIG_PPC_64K_PAGES */
552 static void __init htab_scan_page_sizes(void)
556 /* se the invalid penc to -1 */
557 mmu_psize_set_default_penc();
559 /* Default to 4K pages only */
560 memcpy(mmu_psize_defs, mmu_psize_defaults_old,
561 sizeof(mmu_psize_defaults_old));
564 * Try to find the available page sizes in the device-tree
566 rc = of_scan_flat_dt(htab_dt_scan_page_sizes, NULL);
567 if (rc == 0 && early_mmu_has_feature(MMU_FTR_16M_PAGE)) {
569 * Nothing in the device-tree, but the CPU supports 16M pages,
570 * so let's fallback on a known size list for 16M capable CPUs.
572 memcpy(mmu_psize_defs, mmu_psize_defaults_gp,
573 sizeof(mmu_psize_defaults_gp));
576 #ifdef CONFIG_HUGETLB_PAGE
577 /* Reserve 16G huge page memory sections for huge pages */
578 of_scan_flat_dt(htab_dt_scan_hugepage_blocks, NULL);
579 #endif /* CONFIG_HUGETLB_PAGE */
583 * Fill in the hpte_page_sizes[] array.
584 * We go through the mmu_psize_defs[] array looking for all the
585 * supported base/actual page size combinations. Each combination
586 * has a unique pagesize encoding (penc) value in the low bits of
587 * the LP field of the HPTE. For actual page sizes less than 1MB,
588 * some of the upper LP bits are used for RPN bits, meaning that
589 * we need to fill in several entries in hpte_page_sizes[].
591 * In diagrammatic form, with r = RPN bits and z = page size bits:
592 * PTE LP actual page size
599 * The zzzz bits are implementation-specific but are chosen so that
600 * no encoding for a larger page size uses the same value in its
601 * low-order N bits as the encoding for the 2^(12+N) byte page size
604 static void init_hpte_page_sizes(void)
607 long int shift, penc;
609 for (bp = 0; bp < MMU_PAGE_COUNT; ++bp) {
610 if (!mmu_psize_defs[bp].shift)
611 continue; /* not a supported page size */
612 for (ap = bp; ap < MMU_PAGE_COUNT; ++ap) {
613 penc = mmu_psize_defs[bp].penc[ap];
616 shift = mmu_psize_defs[ap].shift - LP_SHIFT;
618 continue; /* should never happen */
620 * For page sizes less than 1MB, this loop
621 * replicates the entry for all possible values
624 while (penc < (1 << LP_BITS)) {
625 hpte_page_sizes[penc] = (ap << 4) | bp;
632 static void __init htab_init_page_sizes(void)
634 init_hpte_page_sizes();
636 if (!debug_pagealloc_enabled()) {
638 * Pick a size for the linear mapping. Currently, we only
639 * support 16M, 1M and 4K which is the default
641 if (mmu_psize_defs[MMU_PAGE_16M].shift)
642 mmu_linear_psize = MMU_PAGE_16M;
643 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
644 mmu_linear_psize = MMU_PAGE_1M;
647 #ifdef CONFIG_PPC_64K_PAGES
649 * Pick a size for the ordinary pages. Default is 4K, we support
650 * 64K for user mappings and vmalloc if supported by the processor.
651 * We only use 64k for ioremap if the processor
652 * (and firmware) support cache-inhibited large pages.
653 * If not, we use 4k and set mmu_ci_restrictions so that
654 * hash_page knows to switch processes that use cache-inhibited
655 * mappings to 4k pages.
657 if (mmu_psize_defs[MMU_PAGE_64K].shift) {
658 mmu_virtual_psize = MMU_PAGE_64K;
659 mmu_vmalloc_psize = MMU_PAGE_64K;
660 if (mmu_linear_psize == MMU_PAGE_4K)
661 mmu_linear_psize = MMU_PAGE_64K;
662 if (mmu_has_feature(MMU_FTR_CI_LARGE_PAGE)) {
664 * When running on pSeries using 64k pages for ioremap
665 * would stop us accessing the HEA ethernet. So if we
666 * have the chance of ever seeing one, stay at 4k.
668 if (!might_have_hea())
669 mmu_io_psize = MMU_PAGE_64K;
671 mmu_ci_restrictions = 1;
673 #endif /* CONFIG_PPC_64K_PAGES */
675 #ifdef CONFIG_SPARSEMEM_VMEMMAP
676 /* We try to use 16M pages for vmemmap if that is supported
677 * and we have at least 1G of RAM at boot
679 if (mmu_psize_defs[MMU_PAGE_16M].shift &&
680 memblock_phys_mem_size() >= 0x40000000)
681 mmu_vmemmap_psize = MMU_PAGE_16M;
682 else if (mmu_psize_defs[MMU_PAGE_64K].shift)
683 mmu_vmemmap_psize = MMU_PAGE_64K;
685 mmu_vmemmap_psize = MMU_PAGE_4K;
686 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
688 printk(KERN_DEBUG "Page orders: linear mapping = %d, "
689 "virtual = %d, io = %d"
690 #ifdef CONFIG_SPARSEMEM_VMEMMAP
694 mmu_psize_defs[mmu_linear_psize].shift,
695 mmu_psize_defs[mmu_virtual_psize].shift,
696 mmu_psize_defs[mmu_io_psize].shift
697 #ifdef CONFIG_SPARSEMEM_VMEMMAP
698 ,mmu_psize_defs[mmu_vmemmap_psize].shift
703 static int __init htab_dt_scan_pftsize(unsigned long node,
704 const char *uname, int depth,
707 const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
710 /* We are scanning "cpu" nodes only */
711 if (type == NULL || strcmp(type, "cpu") != 0)
714 prop = of_get_flat_dt_prop(node, "ibm,pft-size", NULL);
716 /* pft_size[0] is the NUMA CEC cookie */
717 ppc64_pft_size = be32_to_cpu(prop[1]);
723 unsigned htab_shift_for_mem_size(unsigned long mem_size)
725 unsigned memshift = __ilog2(mem_size);
726 unsigned pshift = mmu_psize_defs[mmu_virtual_psize].shift;
729 /* round mem_size up to next power of 2 */
730 if ((1UL << memshift) < mem_size)
733 /* aim for 2 pages / pteg */
734 pteg_shift = memshift - (pshift + 1);
737 * 2^11 PTEGS of 128 bytes each, ie. 2^18 bytes is the minimum htab
738 * size permitted by the architecture.
740 return max(pteg_shift + 7, 18U);
743 static unsigned long __init htab_get_table_size(void)
745 /* If hash size isn't already provided by the platform, we try to
746 * retrieve it from the device-tree. If it's not there neither, we
747 * calculate it now based on the total RAM size
749 if (ppc64_pft_size == 0)
750 of_scan_flat_dt(htab_dt_scan_pftsize, NULL);
752 return 1UL << ppc64_pft_size;
754 return 1UL << htab_shift_for_mem_size(memblock_phys_mem_size());
757 #ifdef CONFIG_MEMORY_HOTPLUG
758 int hash__create_section_mapping(unsigned long start, unsigned long end)
760 int rc = htab_bolt_mapping(start, end, __pa(start),
761 pgprot_val(PAGE_KERNEL), mmu_linear_psize,
765 int rc2 = htab_remove_mapping(start, end, mmu_linear_psize,
767 BUG_ON(rc2 && (rc2 != -ENOENT));
772 int hash__remove_section_mapping(unsigned long start, unsigned long end)
774 int rc = htab_remove_mapping(start, end, mmu_linear_psize,
779 #endif /* CONFIG_MEMORY_HOTPLUG */
781 static void update_hid_for_hash(void)
784 unsigned long rb = 3UL << PPC_BITLSHIFT(53); /* IS = 3 */
786 asm volatile("ptesync": : :"memory");
787 /* prs = 0, ric = 2, rs = 0, r = 1 is = 3 */
788 asm volatile(PPC_TLBIE_5(%0, %4, %3, %2, %1)
789 : : "r"(rb), "i"(0), "i"(0), "i"(2), "r"(0) : "memory");
790 asm volatile("eieio; tlbsync; ptesync; isync; slbia": : :"memory");
794 hid0 = mfspr(SPRN_HID0);
795 hid0 &= ~HID0_POWER9_RADIX;
796 mtspr(SPRN_HID0, hid0);
797 asm volatile("isync": : :"memory");
799 /* Wait for it to happen */
800 while ((mfspr(SPRN_HID0) & HID0_POWER9_RADIX))
804 static void __init hash_init_partition_table(phys_addr_t hash_table,
805 unsigned long htab_size)
807 unsigned long ps_field;
808 unsigned long patb_size = 1UL << PATB_SIZE_SHIFT;
811 * slb llp encoding for the page size used in VPM real mode.
812 * We can ignore that for lpid 0
815 htab_size = __ilog2(htab_size) - 18;
817 BUILD_BUG_ON_MSG((PATB_SIZE_SHIFT > 24), "Partition table size too large.");
818 partition_tb = __va(memblock_alloc_base(patb_size, patb_size,
819 MEMBLOCK_ALLOC_ANYWHERE));
821 /* Initialize the Partition Table with no entries */
822 memset((void *)partition_tb, 0, patb_size);
823 partition_tb->patb0 = cpu_to_be64(ps_field | hash_table | htab_size);
825 * FIXME!! This should be done via update_partition table
826 * For now UPRT is 0 for us.
828 partition_tb->patb1 = 0;
829 pr_info("Partition table %p\n", partition_tb);
830 if (cpu_has_feature(CPU_FTR_POWER9_DD1))
831 update_hid_for_hash();
833 * update partition table control register,
836 mtspr(SPRN_PTCR, __pa(partition_tb) | (PATB_SIZE_SHIFT - 12));
840 static void __init htab_initialize(void)
843 unsigned long pteg_count;
845 unsigned long base = 0, size = 0;
846 struct memblock_region *reg;
848 DBG(" -> htab_initialize()\n");
850 if (mmu_has_feature(MMU_FTR_1T_SEGMENT)) {
851 mmu_kernel_ssize = MMU_SEGSIZE_1T;
852 mmu_highuser_ssize = MMU_SEGSIZE_1T;
853 printk(KERN_INFO "Using 1TB segments\n");
857 * Calculate the required size of the htab. We want the number of
858 * PTEGs to equal one half the number of real pages.
860 htab_size_bytes = htab_get_table_size();
861 pteg_count = htab_size_bytes >> 7;
863 htab_hash_mask = pteg_count - 1;
865 if (firmware_has_feature(FW_FEATURE_LPAR) ||
866 firmware_has_feature(FW_FEATURE_PS3_LV1)) {
867 /* Using a hypervisor which owns the htab */
870 #ifdef CONFIG_FA_DUMP
872 * If firmware assisted dump is active firmware preserves
873 * the contents of htab along with entire partition memory.
874 * Clear the htab if firmware assisted dump is active so
875 * that we dont end up using old mappings.
877 if (is_fadump_active() && mmu_hash_ops.hpte_clear_all)
878 mmu_hash_ops.hpte_clear_all();
881 unsigned long limit = MEMBLOCK_ALLOC_ANYWHERE;
883 #ifdef CONFIG_PPC_CELL
885 * Cell may require the hash table down low when using the
886 * Axon IOMMU in order to fit the dynamic region over it, see
887 * comments in cell/iommu.c
889 if (fdt_subnode_offset(initial_boot_params, 0, "axon") > 0) {
891 pr_info("Hash table forced below 2G for Axon IOMMU\n");
893 #endif /* CONFIG_PPC_CELL */
895 table = memblock_alloc_base(htab_size_bytes, htab_size_bytes,
898 DBG("Hash table allocated at %lx, size: %lx\n", table,
901 htab_address = __va(table);
903 /* htab absolute addr + encoded htabsize */
904 _SDR1 = table + __ilog2(htab_size_bytes) - 18;
906 /* Initialize the HPT with no entries */
907 memset((void *)table, 0, htab_size_bytes);
909 if (!cpu_has_feature(CPU_FTR_ARCH_300))
911 mtspr(SPRN_SDR1, _SDR1);
913 hash_init_partition_table(table, htab_size_bytes);
916 prot = pgprot_val(PAGE_KERNEL);
918 #ifdef CONFIG_DEBUG_PAGEALLOC
919 if (debug_pagealloc_enabled()) {
920 linear_map_hash_count = memblock_end_of_DRAM() >> PAGE_SHIFT;
921 linear_map_hash_slots = __va(memblock_alloc_base(
922 linear_map_hash_count, 1, ppc64_rma_size));
923 memset(linear_map_hash_slots, 0, linear_map_hash_count);
925 #endif /* CONFIG_DEBUG_PAGEALLOC */
927 /* On U3 based machines, we need to reserve the DART area and
928 * _NOT_ map it to avoid cache paradoxes as it's remapped non
932 /* create bolted the linear mapping in the hash table */
933 for_each_memblock(memory, reg) {
934 base = (unsigned long)__va(reg->base);
937 DBG("creating mapping for region: %lx..%lx (prot: %lx)\n",
940 BUG_ON(htab_bolt_mapping(base, base + size, __pa(base),
941 prot, mmu_linear_psize, mmu_kernel_ssize));
943 memblock_set_current_limit(MEMBLOCK_ALLOC_ANYWHERE);
946 * If we have a memory_limit and we've allocated TCEs then we need to
947 * explicitly map the TCE area at the top of RAM. We also cope with the
948 * case that the TCEs start below memory_limit.
949 * tce_alloc_start/end are 16MB aligned so the mapping should work
950 * for either 4K or 16MB pages.
952 if (tce_alloc_start) {
953 tce_alloc_start = (unsigned long)__va(tce_alloc_start);
954 tce_alloc_end = (unsigned long)__va(tce_alloc_end);
956 if (base + size >= tce_alloc_start)
957 tce_alloc_start = base + size + 1;
959 BUG_ON(htab_bolt_mapping(tce_alloc_start, tce_alloc_end,
960 __pa(tce_alloc_start), prot,
961 mmu_linear_psize, mmu_kernel_ssize));
965 DBG(" <- htab_initialize()\n");
970 void __init hash__early_init_devtree(void)
972 /* Initialize segment sizes */
973 of_scan_flat_dt(htab_dt_scan_seg_sizes, NULL);
975 /* Initialize page sizes */
976 htab_scan_page_sizes();
979 void __init hash__early_init_mmu(void)
981 htab_init_page_sizes();
984 * initialize page table size
986 __pte_frag_nr = H_PTE_FRAG_NR;
987 __pte_frag_size_shift = H_PTE_FRAG_SIZE_SHIFT;
989 __pte_index_size = H_PTE_INDEX_SIZE;
990 __pmd_index_size = H_PMD_INDEX_SIZE;
991 __pud_index_size = H_PUD_INDEX_SIZE;
992 __pgd_index_size = H_PGD_INDEX_SIZE;
993 __pmd_cache_index = H_PMD_CACHE_INDEX;
994 __pte_table_size = H_PTE_TABLE_SIZE;
995 __pmd_table_size = H_PMD_TABLE_SIZE;
996 __pud_table_size = H_PUD_TABLE_SIZE;
997 __pgd_table_size = H_PGD_TABLE_SIZE;
999 * 4k use hugepd format, so for hash set then to
1006 __kernel_virt_start = H_KERN_VIRT_START;
1007 __kernel_virt_size = H_KERN_VIRT_SIZE;
1008 __vmalloc_start = H_VMALLOC_START;
1009 __vmalloc_end = H_VMALLOC_END;
1010 vmemmap = (struct page *)H_VMEMMAP_BASE;
1011 ioremap_bot = IOREMAP_BASE;
1014 pci_io_base = ISA_IO_BASE;
1017 /* Select appropriate backend */
1018 if (firmware_has_feature(FW_FEATURE_PS3_LV1))
1019 ps3_early_mm_init();
1020 else if (firmware_has_feature(FW_FEATURE_LPAR))
1021 hpte_init_pseries();
1022 else if (IS_ENABLED(CONFIG_PPC_NATIVE))
1025 if (!mmu_hash_ops.hpte_insert)
1026 panic("hash__early_init_mmu: No MMU hash ops defined!\n");
1028 /* Initialize the MMU Hash table and create the linear mapping
1029 * of memory. Has to be done before SLB initialization as this is
1030 * currently where the page size encoding is obtained.
1034 pr_info("Initializing hash mmu with SLB\n");
1035 /* Initialize SLB management */
1040 void hash__early_init_mmu_secondary(void)
1042 /* Initialize hash table for that CPU */
1043 if (!firmware_has_feature(FW_FEATURE_LPAR)) {
1045 if (cpu_has_feature(CPU_FTR_POWER9_DD1))
1046 update_hid_for_hash();
1048 if (!cpu_has_feature(CPU_FTR_ARCH_300))
1049 mtspr(SPRN_SDR1, _SDR1);
1052 __pa(partition_tb) | (PATB_SIZE_SHIFT - 12));
1054 /* Initialize SLB */
1057 #endif /* CONFIG_SMP */
1060 * Called by asm hashtable.S for doing lazy icache flush
1062 unsigned int hash_page_do_lazy_icache(unsigned int pp, pte_t pte, int trap)
1066 if (!pfn_valid(pte_pfn(pte)))
1069 page = pte_page(pte);
1072 if (!test_bit(PG_arch_1, &page->flags) && !PageReserved(page)) {
1073 if (trap == 0x400) {
1074 flush_dcache_icache_page(page);
1075 set_bit(PG_arch_1, &page->flags);
1082 #ifdef CONFIG_PPC_MM_SLICES
1083 static unsigned int get_paca_psize(unsigned long addr)
1086 unsigned char *hpsizes;
1087 unsigned long index, mask_index;
1089 if (addr < SLICE_LOW_TOP) {
1090 lpsizes = get_paca()->mm_ctx_low_slices_psize;
1091 index = GET_LOW_SLICE_INDEX(addr);
1092 return (lpsizes >> (index * 4)) & 0xF;
1094 hpsizes = get_paca()->mm_ctx_high_slices_psize;
1095 index = GET_HIGH_SLICE_INDEX(addr);
1096 mask_index = index & 0x1;
1097 return (hpsizes[index >> 1] >> (mask_index * 4)) & 0xF;
1101 unsigned int get_paca_psize(unsigned long addr)
1103 return get_paca()->mm_ctx_user_psize;
1108 * Demote a segment to using 4k pages.
1109 * For now this makes the whole process use 4k pages.
1111 #ifdef CONFIG_PPC_64K_PAGES
1112 void demote_segment_4k(struct mm_struct *mm, unsigned long addr)
1114 if (get_slice_psize(mm, addr) == MMU_PAGE_4K)
1116 slice_set_range_psize(mm, addr, 1, MMU_PAGE_4K);
1117 copro_flush_all_slbs(mm);
1118 if ((get_paca_psize(addr) != MMU_PAGE_4K) && (current->mm == mm)) {
1120 copy_mm_to_paca(&mm->context);
1121 slb_flush_and_rebolt();
1124 #endif /* CONFIG_PPC_64K_PAGES */
1126 #ifdef CONFIG_PPC_SUBPAGE_PROT
1128 * This looks up a 2-bit protection code for a 4k subpage of a 64k page.
1129 * Userspace sets the subpage permissions using the subpage_prot system call.
1131 * Result is 0: full permissions, _PAGE_RW: read-only,
1132 * _PAGE_RWX: no access.
1134 static int subpage_protection(struct mm_struct *mm, unsigned long ea)
1136 struct subpage_prot_table *spt = &mm->context.spt;
1140 if (ea >= spt->maxaddr)
1142 if (ea < 0x100000000UL) {
1143 /* addresses below 4GB use spt->low_prot */
1144 sbpm = spt->low_prot;
1146 sbpm = spt->protptrs[ea >> SBP_L3_SHIFT];
1150 sbpp = sbpm[(ea >> SBP_L2_SHIFT) & (SBP_L2_COUNT - 1)];
1153 spp = sbpp[(ea >> PAGE_SHIFT) & (SBP_L1_COUNT - 1)];
1155 /* extract 2-bit bitfield for this 4k subpage */
1156 spp >>= 30 - 2 * ((ea >> 12) & 0xf);
1159 * 0 -> full premission
1162 * We return the flag that need to be cleared.
1164 spp = ((spp & 2) ? _PAGE_RWX : 0) | ((spp & 1) ? _PAGE_WRITE : 0);
1168 #else /* CONFIG_PPC_SUBPAGE_PROT */
1169 static inline int subpage_protection(struct mm_struct *mm, unsigned long ea)
1175 void hash_failure_debug(unsigned long ea, unsigned long access,
1176 unsigned long vsid, unsigned long trap,
1177 int ssize, int psize, int lpsize, unsigned long pte)
1179 if (!printk_ratelimit())
1181 pr_info("mm: Hashing failure ! EA=0x%lx access=0x%lx current=%s\n",
1182 ea, access, current->comm);
1183 pr_info(" trap=0x%lx vsid=0x%lx ssize=%d base psize=%d psize %d pte=0x%lx\n",
1184 trap, vsid, ssize, psize, lpsize, pte);
1187 static void check_paca_psize(unsigned long ea, struct mm_struct *mm,
1188 int psize, bool user_region)
1191 if (psize != get_paca_psize(ea)) {
1192 copy_mm_to_paca(&mm->context);
1193 slb_flush_and_rebolt();
1195 } else if (get_paca()->vmalloc_sllp !=
1196 mmu_psize_defs[mmu_vmalloc_psize].sllp) {
1197 get_paca()->vmalloc_sllp =
1198 mmu_psize_defs[mmu_vmalloc_psize].sllp;
1199 slb_vmalloc_update();
1205 * 1 - normal page fault
1206 * -1 - critical hash insertion error
1207 * -2 - access not permitted by subpage protection mechanism
1209 int hash_page_mm(struct mm_struct *mm, unsigned long ea,
1210 unsigned long access, unsigned long trap,
1211 unsigned long flags)
1214 enum ctx_state prev_state = exception_enter();
1219 const struct cpumask *tmp;
1220 int rc, user_region = 0;
1223 DBG_LOW("hash_page(ea=%016lx, access=%lx, trap=%lx\n",
1225 trace_hash_fault(ea, access, trap);
1227 /* Get region & vsid */
1228 switch (REGION_ID(ea)) {
1229 case USER_REGION_ID:
1232 DBG_LOW(" user region with no mm !\n");
1236 psize = get_slice_psize(mm, ea);
1237 ssize = user_segment_size(ea);
1238 vsid = get_vsid(mm->context.id, ea, ssize);
1240 case VMALLOC_REGION_ID:
1241 vsid = get_kernel_vsid(ea, mmu_kernel_ssize);
1242 if (ea < VMALLOC_END)
1243 psize = mmu_vmalloc_psize;
1245 psize = mmu_io_psize;
1246 ssize = mmu_kernel_ssize;
1249 /* Not a valid range
1250 * Send the problem up to do_page_fault
1255 DBG_LOW(" mm=%p, mm->pgdir=%p, vsid=%016lx\n", mm, mm->pgd, vsid);
1259 DBG_LOW("Bad address!\n");
1265 if (pgdir == NULL) {
1270 /* Check CPU locality */
1271 tmp = cpumask_of(smp_processor_id());
1272 if (user_region && cpumask_equal(mm_cpumask(mm), tmp))
1273 flags |= HPTE_LOCAL_UPDATE;
1275 #ifndef CONFIG_PPC_64K_PAGES
1276 /* If we use 4K pages and our psize is not 4K, then we might
1277 * be hitting a special driver mapping, and need to align the
1278 * address before we fetch the PTE.
1280 * It could also be a hugepage mapping, in which case this is
1281 * not necessary, but it's not harmful, either.
1283 if (psize != MMU_PAGE_4K)
1284 ea &= ~((1ul << mmu_psize_defs[psize].shift) - 1);
1285 #endif /* CONFIG_PPC_64K_PAGES */
1287 /* Get PTE and page size from page tables */
1288 ptep = __find_linux_pte_or_hugepte(pgdir, ea, &is_thp, &hugeshift);
1289 if (ptep == NULL || !pte_present(*ptep)) {
1290 DBG_LOW(" no PTE !\n");
1295 /* Add _PAGE_PRESENT to the required access perm */
1296 access |= _PAGE_PRESENT;
1298 /* Pre-check access permissions (will be re-checked atomically
1299 * in __hash_page_XX but this pre-check is a fast path
1301 if (!check_pte_access(access, pte_val(*ptep))) {
1302 DBG_LOW(" no access !\n");
1309 rc = __hash_page_thp(ea, access, vsid, (pmd_t *)ptep,
1310 trap, flags, ssize, psize);
1311 #ifdef CONFIG_HUGETLB_PAGE
1313 rc = __hash_page_huge(ea, access, vsid, ptep, trap,
1314 flags, ssize, hugeshift, psize);
1318 * if we have hugeshift, and is not transhuge with
1319 * hugetlb disabled, something is really wrong.
1325 if (current->mm == mm)
1326 check_paca_psize(ea, mm, psize, user_region);
1331 #ifndef CONFIG_PPC_64K_PAGES
1332 DBG_LOW(" i-pte: %016lx\n", pte_val(*ptep));
1334 DBG_LOW(" i-pte: %016lx %016lx\n", pte_val(*ptep),
1335 pte_val(*(ptep + PTRS_PER_PTE)));
1337 /* Do actual hashing */
1338 #ifdef CONFIG_PPC_64K_PAGES
1339 /* If H_PAGE_4K_PFN is set, make sure this is a 4k segment */
1340 if ((pte_val(*ptep) & H_PAGE_4K_PFN) && psize == MMU_PAGE_64K) {
1341 demote_segment_4k(mm, ea);
1342 psize = MMU_PAGE_4K;
1345 /* If this PTE is non-cacheable and we have restrictions on
1346 * using non cacheable large pages, then we switch to 4k
1348 if (mmu_ci_restrictions && psize == MMU_PAGE_64K && pte_ci(*ptep)) {
1350 demote_segment_4k(mm, ea);
1351 psize = MMU_PAGE_4K;
1352 } else if (ea < VMALLOC_END) {
1354 * some driver did a non-cacheable mapping
1355 * in vmalloc space, so switch vmalloc
1358 printk(KERN_ALERT "Reducing vmalloc segment "
1359 "to 4kB pages because of "
1360 "non-cacheable mapping\n");
1361 psize = mmu_vmalloc_psize = MMU_PAGE_4K;
1362 copro_flush_all_slbs(mm);
1366 #endif /* CONFIG_PPC_64K_PAGES */
1368 if (current->mm == mm)
1369 check_paca_psize(ea, mm, psize, user_region);
1371 #ifdef CONFIG_PPC_64K_PAGES
1372 if (psize == MMU_PAGE_64K)
1373 rc = __hash_page_64K(ea, access, vsid, ptep, trap,
1376 #endif /* CONFIG_PPC_64K_PAGES */
1378 int spp = subpage_protection(mm, ea);
1382 rc = __hash_page_4K(ea, access, vsid, ptep, trap,
1386 /* Dump some info in case of hash insertion failure, they should
1387 * never happen so it is really useful to know if/when they do
1390 hash_failure_debug(ea, access, vsid, trap, ssize, psize,
1391 psize, pte_val(*ptep));
1392 #ifndef CONFIG_PPC_64K_PAGES
1393 DBG_LOW(" o-pte: %016lx\n", pte_val(*ptep));
1395 DBG_LOW(" o-pte: %016lx %016lx\n", pte_val(*ptep),
1396 pte_val(*(ptep + PTRS_PER_PTE)));
1398 DBG_LOW(" -> rc=%d\n", rc);
1401 exception_exit(prev_state);
1404 EXPORT_SYMBOL_GPL(hash_page_mm);
1406 int hash_page(unsigned long ea, unsigned long access, unsigned long trap,
1407 unsigned long dsisr)
1409 unsigned long flags = 0;
1410 struct mm_struct *mm = current->mm;
1412 if (REGION_ID(ea) == VMALLOC_REGION_ID)
1415 if (dsisr & DSISR_NOHPTE)
1416 flags |= HPTE_NOHPTE_UPDATE;
1418 return hash_page_mm(mm, ea, access, trap, flags);
1420 EXPORT_SYMBOL_GPL(hash_page);
1422 int __hash_page(unsigned long ea, unsigned long msr, unsigned long trap,
1423 unsigned long dsisr)
1425 unsigned long access = _PAGE_PRESENT | _PAGE_READ;
1426 unsigned long flags = 0;
1427 struct mm_struct *mm = current->mm;
1429 if (REGION_ID(ea) == VMALLOC_REGION_ID)
1432 if (dsisr & DSISR_NOHPTE)
1433 flags |= HPTE_NOHPTE_UPDATE;
1435 if (dsisr & DSISR_ISSTORE)
1436 access |= _PAGE_WRITE;
1438 * We set _PAGE_PRIVILEGED only when
1439 * kernel mode access kernel space.
1441 * _PAGE_PRIVILEGED is NOT set
1442 * 1) when kernel mode access user space
1443 * 2) user space access kernel space.
1445 access |= _PAGE_PRIVILEGED;
1446 if ((msr & MSR_PR) || (REGION_ID(ea) == USER_REGION_ID))
1447 access &= ~_PAGE_PRIVILEGED;
1450 access |= _PAGE_EXEC;
1452 return hash_page_mm(mm, ea, access, trap, flags);
1455 #ifdef CONFIG_PPC_MM_SLICES
1456 static bool should_hash_preload(struct mm_struct *mm, unsigned long ea)
1458 int psize = get_slice_psize(mm, ea);
1460 /* We only prefault standard pages for now */
1461 if (unlikely(psize != mm->context.user_psize))
1465 * Don't prefault if subpage protection is enabled for the EA.
1467 if (unlikely((psize == MMU_PAGE_4K) && subpage_protection(mm, ea)))
1473 static bool should_hash_preload(struct mm_struct *mm, unsigned long ea)
1479 void hash_preload(struct mm_struct *mm, unsigned long ea,
1480 unsigned long access, unsigned long trap)
1486 unsigned long flags;
1487 int rc, ssize, update_flags = 0;
1489 BUG_ON(REGION_ID(ea) != USER_REGION_ID);
1491 if (!should_hash_preload(mm, ea))
1494 DBG_LOW("hash_preload(mm=%p, mm->pgdir=%p, ea=%016lx, access=%lx,"
1495 " trap=%lx\n", mm, mm->pgd, ea, access, trap);
1497 /* Get Linux PTE if available */
1503 ssize = user_segment_size(ea);
1504 vsid = get_vsid(mm->context.id, ea, ssize);
1508 * Hash doesn't like irqs. Walking linux page table with irq disabled
1509 * saves us from holding multiple locks.
1511 local_irq_save(flags);
1514 * THP pages use update_mmu_cache_pmd. We don't do
1515 * hash preload there. Hence can ignore THP here
1517 ptep = find_linux_pte_or_hugepte(pgdir, ea, NULL, &hugepage_shift);
1521 WARN_ON(hugepage_shift);
1522 #ifdef CONFIG_PPC_64K_PAGES
1523 /* If either H_PAGE_4K_PFN or cache inhibited is set (and we are on
1524 * a 64K kernel), then we don't preload, hash_page() will take
1525 * care of it once we actually try to access the page.
1526 * That way we don't have to duplicate all of the logic for segment
1527 * page size demotion here
1529 if ((pte_val(*ptep) & H_PAGE_4K_PFN) || pte_ci(*ptep))
1531 #endif /* CONFIG_PPC_64K_PAGES */
1533 /* Is that local to this CPU ? */
1534 if (cpumask_equal(mm_cpumask(mm), cpumask_of(smp_processor_id())))
1535 update_flags |= HPTE_LOCAL_UPDATE;
1538 #ifdef CONFIG_PPC_64K_PAGES
1539 if (mm->context.user_psize == MMU_PAGE_64K)
1540 rc = __hash_page_64K(ea, access, vsid, ptep, trap,
1541 update_flags, ssize);
1543 #endif /* CONFIG_PPC_64K_PAGES */
1544 rc = __hash_page_4K(ea, access, vsid, ptep, trap, update_flags,
1545 ssize, subpage_protection(mm, ea));
1547 /* Dump some info in case of hash insertion failure, they should
1548 * never happen so it is really useful to know if/when they do
1551 hash_failure_debug(ea, access, vsid, trap, ssize,
1552 mm->context.user_psize,
1553 mm->context.user_psize,
1556 local_irq_restore(flags);
1559 #ifdef CONFIG_PPC_TRANSACTIONAL_MEM
1560 static inline void tm_flush_hash_page(int local)
1563 * Transactions are not aborted by tlbiel, only tlbie. Without, syncing a
1564 * page back to a block device w/PIO could pick up transactional data
1565 * (bad!) so we force an abort here. Before the sync the page will be
1566 * made read-only, which will flush_hash_page. BIG ISSUE here: if the
1567 * kernel uses a page from userspace without unmapping it first, it may
1568 * see the speculated version.
1570 if (local && cpu_has_feature(CPU_FTR_TM) && current->thread.regs &&
1571 MSR_TM_ACTIVE(current->thread.regs->msr)) {
1573 tm_abort(TM_CAUSE_TLBI);
1577 static inline void tm_flush_hash_page(int local)
1582 /* WARNING: This is called from hash_low_64.S, if you change this prototype,
1583 * do not forget to update the assembly call site !
1585 void flush_hash_page(unsigned long vpn, real_pte_t pte, int psize, int ssize,
1586 unsigned long flags)
1588 unsigned long hash, index, shift, hidx, slot;
1589 int local = flags & HPTE_LOCAL_UPDATE;
1591 DBG_LOW("flush_hash_page(vpn=%016lx)\n", vpn);
1592 pte_iterate_hashed_subpages(pte, psize, vpn, index, shift) {
1593 hash = hpt_hash(vpn, shift, ssize);
1594 hidx = __rpte_to_hidx(pte, index);
1595 if (hidx & _PTEIDX_SECONDARY)
1597 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1598 slot += hidx & _PTEIDX_GROUP_IX;
1599 DBG_LOW(" sub %ld: hash=%lx, hidx=%lx\n", index, slot, hidx);
1601 * We use same base page size and actual psize, because we don't
1602 * use these functions for hugepage
1604 mmu_hash_ops.hpte_invalidate(slot, vpn, psize, psize,
1606 } pte_iterate_hashed_end();
1608 tm_flush_hash_page(local);
1611 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1612 void flush_hash_hugepage(unsigned long vsid, unsigned long addr,
1613 pmd_t *pmdp, unsigned int psize, int ssize,
1614 unsigned long flags)
1616 int i, max_hpte_count, valid;
1617 unsigned long s_addr;
1618 unsigned char *hpte_slot_array;
1619 unsigned long hidx, shift, vpn, hash, slot;
1620 int local = flags & HPTE_LOCAL_UPDATE;
1622 s_addr = addr & HPAGE_PMD_MASK;
1623 hpte_slot_array = get_hpte_slot_array(pmdp);
1625 * IF we try to do a HUGE PTE update after a withdraw is done.
1626 * we will find the below NULL. This happens when we do
1627 * split_huge_page_pmd
1629 if (!hpte_slot_array)
1632 if (mmu_hash_ops.hugepage_invalidate) {
1633 mmu_hash_ops.hugepage_invalidate(vsid, s_addr, hpte_slot_array,
1634 psize, ssize, local);
1638 * No bluk hpte removal support, invalidate each entry
1640 shift = mmu_psize_defs[psize].shift;
1641 max_hpte_count = HPAGE_PMD_SIZE >> shift;
1642 for (i = 0; i < max_hpte_count; i++) {
1644 * 8 bits per each hpte entries
1645 * 000| [ secondary group (one bit) | hidx (3 bits) | valid bit]
1647 valid = hpte_valid(hpte_slot_array, i);
1650 hidx = hpte_hash_index(hpte_slot_array, i);
1653 addr = s_addr + (i * (1ul << shift));
1654 vpn = hpt_vpn(addr, vsid, ssize);
1655 hash = hpt_hash(vpn, shift, ssize);
1656 if (hidx & _PTEIDX_SECONDARY)
1659 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1660 slot += hidx & _PTEIDX_GROUP_IX;
1661 mmu_hash_ops.hpte_invalidate(slot, vpn, psize,
1662 MMU_PAGE_16M, ssize, local);
1665 tm_flush_hash_page(local);
1667 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1669 void flush_hash_range(unsigned long number, int local)
1671 if (mmu_hash_ops.flush_hash_range)
1672 mmu_hash_ops.flush_hash_range(number, local);
1675 struct ppc64_tlb_batch *batch =
1676 this_cpu_ptr(&ppc64_tlb_batch);
1678 for (i = 0; i < number; i++)
1679 flush_hash_page(batch->vpn[i], batch->pte[i],
1680 batch->psize, batch->ssize, local);
1685 * low_hash_fault is called when we the low level hash code failed
1686 * to instert a PTE due to an hypervisor error
1688 void low_hash_fault(struct pt_regs *regs, unsigned long address, int rc)
1690 enum ctx_state prev_state = exception_enter();
1692 if (user_mode(regs)) {
1693 #ifdef CONFIG_PPC_SUBPAGE_PROT
1695 _exception(SIGSEGV, regs, SEGV_ACCERR, address);
1698 _exception(SIGBUS, regs, BUS_ADRERR, address);
1700 bad_page_fault(regs, address, SIGBUS);
1702 exception_exit(prev_state);
1705 long hpte_insert_repeating(unsigned long hash, unsigned long vpn,
1706 unsigned long pa, unsigned long rflags,
1707 unsigned long vflags, int psize, int ssize)
1709 unsigned long hpte_group;
1713 hpte_group = ((hash & htab_hash_mask) *
1714 HPTES_PER_GROUP) & ~0x7UL;
1716 /* Insert into the hash table, primary slot */
1717 slot = mmu_hash_ops.hpte_insert(hpte_group, vpn, pa, rflags, vflags,
1718 psize, psize, ssize);
1720 /* Primary is full, try the secondary */
1721 if (unlikely(slot == -1)) {
1722 hpte_group = ((~hash & htab_hash_mask) *
1723 HPTES_PER_GROUP) & ~0x7UL;
1724 slot = mmu_hash_ops.hpte_insert(hpte_group, vpn, pa, rflags,
1725 vflags | HPTE_V_SECONDARY,
1726 psize, psize, ssize);
1729 hpte_group = ((hash & htab_hash_mask) *
1730 HPTES_PER_GROUP)&~0x7UL;
1732 mmu_hash_ops.hpte_remove(hpte_group);
1740 #ifdef CONFIG_DEBUG_PAGEALLOC
1741 static void kernel_map_linear_page(unsigned long vaddr, unsigned long lmi)
1744 unsigned long vsid = get_kernel_vsid(vaddr, mmu_kernel_ssize);
1745 unsigned long vpn = hpt_vpn(vaddr, vsid, mmu_kernel_ssize);
1746 unsigned long mode = htab_convert_pte_flags(pgprot_val(PAGE_KERNEL));
1749 hash = hpt_hash(vpn, PAGE_SHIFT, mmu_kernel_ssize);
1751 /* Don't create HPTE entries for bad address */
1755 ret = hpte_insert_repeating(hash, vpn, __pa(vaddr), mode,
1757 mmu_linear_psize, mmu_kernel_ssize);
1760 spin_lock(&linear_map_hash_lock);
1761 BUG_ON(linear_map_hash_slots[lmi] & 0x80);
1762 linear_map_hash_slots[lmi] = ret | 0x80;
1763 spin_unlock(&linear_map_hash_lock);
1766 static void kernel_unmap_linear_page(unsigned long vaddr, unsigned long lmi)
1768 unsigned long hash, hidx, slot;
1769 unsigned long vsid = get_kernel_vsid(vaddr, mmu_kernel_ssize);
1770 unsigned long vpn = hpt_vpn(vaddr, vsid, mmu_kernel_ssize);
1772 hash = hpt_hash(vpn, PAGE_SHIFT, mmu_kernel_ssize);
1773 spin_lock(&linear_map_hash_lock);
1774 BUG_ON(!(linear_map_hash_slots[lmi] & 0x80));
1775 hidx = linear_map_hash_slots[lmi] & 0x7f;
1776 linear_map_hash_slots[lmi] = 0;
1777 spin_unlock(&linear_map_hash_lock);
1778 if (hidx & _PTEIDX_SECONDARY)
1780 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1781 slot += hidx & _PTEIDX_GROUP_IX;
1782 mmu_hash_ops.hpte_invalidate(slot, vpn, mmu_linear_psize,
1784 mmu_kernel_ssize, 0);
1787 void __kernel_map_pages(struct page *page, int numpages, int enable)
1789 unsigned long flags, vaddr, lmi;
1792 local_irq_save(flags);
1793 for (i = 0; i < numpages; i++, page++) {
1794 vaddr = (unsigned long)page_address(page);
1795 lmi = __pa(vaddr) >> PAGE_SHIFT;
1796 if (lmi >= linear_map_hash_count)
1799 kernel_map_linear_page(vaddr, lmi);
1801 kernel_unmap_linear_page(vaddr, lmi);
1803 local_irq_restore(flags);
1805 #endif /* CONFIG_DEBUG_PAGEALLOC */
1807 void hash__setup_initial_memory_limit(phys_addr_t first_memblock_base,
1808 phys_addr_t first_memblock_size)
1810 /* We don't currently support the first MEMBLOCK not mapping 0
1811 * physical on those processors
1813 BUG_ON(first_memblock_base != 0);
1815 /* On LPAR systems, the first entry is our RMA region,
1816 * non-LPAR 64-bit hash MMU systems don't have a limitation
1817 * on real mode access, but using the first entry works well
1818 * enough. We also clamp it to 1G to avoid some funky things
1819 * such as RTAS bugs etc...
1821 ppc64_rma_size = min_t(u64, first_memblock_size, 0x40000000);
1823 /* Finally limit subsequent allocations */
1824 memblock_set_current_limit(ppc64_rma_size);